Discussion: size optimization (-Os) for Apple platforms

[pblazej]

Context

The Apple xcframework is built with -O2 (optimize for speed), while the android_prefixed_stripped target uses optimize_for_size=true (-Os). This is inherited Chromium behavior — the GN default in build/config/compiler/compiler.gni

optimize_for_size = !is_high_end_android && (is_android || is_castos || is_fuchsia)

opts only Android/CastOS/Fuchsia into size optimization ("favoring speed over size for platforms not listed below"), so Apple lands on the -O2 branch of config("optimize"). The build orchestration (webrtc-sdk/webrtc-build: apple/xcframework.sh → COMMON_ARGS) doesn't override it.

There's no technical blocker to building Apple with -Os — android_prefixed_stripped (webrtc-sdk/webrtc-build#50, M144.7559.01) already establishes the pattern. Opening this to discuss aligning Apple with that approach.

Measured impact

LiveKit-prefixed WebRTC, iOS arm64 device slice, M144 (aea24a1…), Release, stripped:

Build	Uncompressed	Δ install	gzip (≈ download)	Δ download
-O2 (current)	11.29 MB	—	5.24 MB	—
-Os (optimize_for_size=true)	9.95 MB	−1.34 MB (−11.9%)	4.62 MB	−0.62 MB (−11.8%)
-Oz (reference only — see below)	9.39 MB	−1.91 MB (−16.9%)	4.26 MB	−0.97 MB (−18.6%)
-Oz + ThinLTO (reference only)	9.23 MB	−2.06 MB (−18.3%)	4.17 MB	−1.07 MB (−20.4%)

-Os is the relevant row — it's the same flag the Android stripped target already ships. The -Oz/ThinLTO rows are an upper bound only; neither is used anywhere in this project today (Android-stripped is -Os; -Oz is Chromium-gated to Fuchsia/cast only).

Notes:

• Savings are almost entirely __TEXT,__text (code); __DATA is unchanged.
• The hot media codecs (libvpx/opus/dav1d/aom) compile under :optimize_max regardless of optimize_for_size, so they stay -O2+. These deltas are the non-codec WebRTC code shrinking — the realtime encode/decode paths are unaffected.

Proposed change

Add optimize_for_size=true to the Apple build's GN args (apple/xcframework.sh COMMON_ARGS in webrtc-sdk/webrtc-build) — a one-line, no-patch change that mirrors android_prefixed_stripped. Chromium's config("optimize") already emits -Os for Apple once the flag is set.

Options for how far to go:

1. -Os on the shipped Apple framework (recommended starting point) — precedented, low-risk (-Os is "size-conscious -O2"), ~1.3 MB install / ~0.6 MB download per arch.
2. A separate apple_prefixed_stripped variant — mirror the Android stripped target (size + codec trims) as an opt-in artifact, leaving the default on -O2.
3. Status quo — keep -O2 if Apple realtime headroom is worth more than ~1.3 MB.

-Oz/ThinLTO are intentionally not proposed here: no precedent in the project, -Oz needs a patch to the GN config, and the extra ~5% isn't worth the added perf risk + maintenance unless there's specific appetite. Easy to revisit separately.

Open questions

• Perf: codecs are protected (optimize_max), but -Os still touches pacing / jitter-buffer / network / signaling code. Has anyone benchmarked -Os WebRTC on Apple for encode/decode latency, CPU, battery?
• Default vs opt-in: apply to the single shipped framework, or ship a separate _stripped (extra publish + larger release matrix)?
• Intentionality: is Apple's speed default deliberate, or unconsidered Chromium inheritance?

Happy to share full bloaty breakdowns and repro scripts.